> On Oct 2, 2015, at 06:44 , Stephen Satchell <l...@satchell.net> wrote: > > On 10/02/2015 12:44 AM, valdis.kletni...@vt.edu wrote: >> On Fri, 02 Oct 2015 02:09:00 -0400, Rob McEwen said: >> >>> Likewise, sub-allocations can come into play, where a hoster is >>> delegated a /48, but then subdivides it for various customers. >> >> So they apply for a /32 and give each customer a /48. >> >> A hoster getting *just* a /48 is about as silly as a hoster >> getting a /32 of IPv4 and NAT'ing their customers. >> > > I agree, for a web hosting operation, getting an allocation smaller than a > /32 doesn't make sense. > > But...now I ask this question: WHY a /48 per customer? I used to be a web > host guy, and the rule was one IPv4 address per co-location customer or > dedicated-server customer -- maybe two -- and shared-IP HTTP for those > customers hosted on "house" servers with multiple sites on them. We had a > couple of shared-hosting server with 64 IPv4 addresses each to support SSL > sites with customer-provided SSL certificates.. > > OLD STYLE > > If a customer wanted more than one IPv4 address, he had to justify it so we > could copy the justification to our ARIN paperwork. A /24 was right out, > because the *only* people requesting that much IPv4 space were spammers. > > The largest legit co-location IPv4 customer allocation, because he had enough > servers in his cage and sufficient justification to warrant it, was a /26 . > Which I SWIPped. Which I treated as a completely separate subnet. Which was > on its own VLAN. Which used separate 10base-T Ethernet interfaces on my edge > routers to provide hard flow control and traffic monitoring. > > THAT WAS THEN, THIS IS NOW > > I can see, in shared hosting, where each customer gets one IPv6 address to > support HTTPS "properly". Each physical server typically hosts 300-400 web > sites comfortably, so assigning a /112 to each of those servers appears to > make sense. This is particularly true now that there is a push for "https > everywhere". > > Web hosting isn't going to be a downstream link for IoT, so the need for > "massive" amounts of IPv6 addressing space is simply not there. >
So there are a number of reasons. First, unless you want to be chasing ND Cache Overflow problems, you put each customer on a small link (/127) to your router and then route at least a /64 to their router if they just have one subnet. If they have more than one, then you certainly want to route them a larger prefix (/48). With virtualization and network virtualization, containers, and the like these days, treating each customer as a separate end-site is just good practice. You’re not going to have any problem explaining that to the RIRs. Owen
